Apppublic

WillHeld/delphi-1e23-demo

sourceHugging Faceupdated 4mo agoView on Hugging Face

0likes

App README

A few-shot demo of `marin-community/delphi-1e23-25Bparams-628Btokens`, the largest run from the Delphi scaling suite.

This is a pure base model — no midtraining, no SFT, no RLHF — so it only does next-token prediction. The Space wraps each tab with the canonical few-shot prompt for the benchmark before sending it to the model:

—MMLU (5-shot) — multiple-choice knowledge questions
—GSM8K (5-shot) — chain-of-thought grade-school math
—HumanEval (3-shot) — Python function completion

Running the Space

The Space targets ZeroGPU, which dynamically allocates half-H200 (70 GB VRAM, large) or full-H200 (141 GB, xlarge) slots. The 25B model in bf16 (~50 GB) fits the default large slot. Generation is wrapped in @spaces.GPU(duration=120); the model is placed on cuda at module level per the ZeroGPU docs (PyTorch CUDA emulation handles this outside the GPU function).

To run locally with a GPU:

bash

pip install -r requirements.txt
python app.py

DELPHI_MODEL_ID overrides the default checkpoint if you want to swap in a different rung of the suite.